unusualevent

50 %
50 %
Information about unusualevent

Published on October 10, 2007

Author: guestfbf1e1

Source: slideshare.net

REAL-TIME DETECTOR FOR UNUSUAL BEHAVIOR Showcase

Highlights Events usual  non-usual Motion and shape based Statistically relevant  irrelevant Alert generation on unusual event Storing events in database

Events

usual  non-usual

Motion and shape based

Statistically relevant  irrelevant

Alert generation on unusual event

Storing events in database

Platform Visualisation: Web browser SZTAKI will provide a communication module that will call the module functions provided by the partners. Software platform: C++, OpenCv, IPP Web technologi Hard ware platform: Pc, laptop (x86 like)

Visualisation: Web browser

SZTAKI will provide a communication module that will call the module functions provided by the partners.

Software platform:

C++, OpenCv, IPP

Web technologi

Hard ware platform:

Pc, laptop (x86 like)

Partners ACV BILKENT UPC SZTAKI Tracking, pedestrian detection Multimodal human actions, HMM 2D Body actions, motion fields Unusual event detection, annotation process, statistical analysis, shadow removing ( f ormerly A RC ) BILKENT

ACV

BILKENT

UPC

SZTAKI

Distribution of work Moving Cam. Static Cam. mosaicing Foreground Detect. shilouettes HMM class. Body model Motion features Periodicity Pedestrian detection Tracking sound classification Unusual event Region alert Sztaki ACV BILKENT UPC

Contribution of ACV Non-parametric clustering of moving objects in difference images Occlusion handling for interacting targets Kernel-based tracking using motion features for multiple targets Video data set and evaluation of the motion detection and tracking performance (Benchmark competition)

Non-parametric clustering of moving objects in difference images

Occlusion handling for interacting targets

Kernel-based tracking using motion features for multiple targets

Video data set and evaluation of the motion detection and tracking performance (Benchmark competition)

Details on the algorithmic modules Human detection by clustering and model-based verification VIDEO Kernel-based human tracking using motion information VIDEO Occlusion handling VIDEO Tracking evaluation (comparison to manual ground truth)

Evaluation video datasets (street scenarios) Sequence Street_01.avi: 720x576 pixels, 8628 frames (tracking ground truth available for 1040 frames) Sequence Street_02.avi: 720x576 pixels, 763 frames (tracking ground truth available for 763 frames)

Contribution of Bilkent Motion and silhoutte based person detector detect motion and moving blocks and observe periodicity in bounding boxes of moving blocks in video. use silhouttes to classify moving objects in video combine the results of periodicity and silhoutte based detectorIn this way, Determine the number of people in the scene. HMM classification (fight, fall or simply walk) Record the sounds and classify the sounds to (car sounds, walking person, and loud screems) Combine the results of 3 and 4 to reach a final decision. BILKENT

Motion and silhoutte based person detector

detect motion and moving blocks and observe periodicity in bounding boxes of moving blocks in video.

use silhouttes to classify moving objects in video

combine the results of periodicity and silhoutte based detectorIn this way,

Determine the number of people in the scene.

HMM classification (fight, fall or simply walk)

Record the sounds and classify the sounds to (car sounds, walking person, and loud screems)

Combine the results of 3 and 4 to reach a final decision.

Human Recognition in Video Utilizes objects’ silhouettes for different poses Silhouettes are extracted using contour tracing Compare silhouette signature functions using wavelet energy signatures BILKENT

Utilizes objects’ silhouettes for different poses

Silhouettes are extracted using contour tracing

Compare silhouette signature functions using wavelet energy signatures

Observation: Walking and falling person Falling Person Detection using Motion Clues ( visual ) T1 T2 BILKENT

Observation: Walking and falling person

Contribution of UPC Foreground detection and automatic features extraction motion history descriptors simple body model Apply the integrated system to different environments crowded scenes in automatic stairs

Foreground detection and automatic features extraction

motion history descriptors

simple body model

Apply the integrated system to different environments

crowded scenes in automatic stairs

Motion Analysis Motion History and Motion Energy descriptors introduced by Bobick et al. in 2D and Canton et al. in 3D allows robust motion analysis MEV MHV

Motion History and Motion Energy descriptors introduced by Bobick et al. in 2D and Canton et al. in 3D allows robust motion analysis

Model Based Analysis Analyzing input data by means of a Human Body Model, allows retrieving information about limbs positions

Analyzing input data by means of a Human Body Model, allows retrieving information about limbs positions

Silhouette analysis for detection of body extremities Scene capture User segmentation CoG computation Creation of the geodesic distance map Contour tracking Creation of the distance/silhouette border position function H-maxima operation on the function Local maxima extraction Morphological skeleton computation and crucial point labeling Pixel position Geodesic Distance

Scene capture

User segmentation

CoG computation

Creation of the geodesic distance map

Contour tracking

Creation of the distance/silhouette border position function

H-maxima operation on the function

Local maxima extraction

Morphological skeleton computation and crucial point labeling

Contribution of SZTAKI Foreground detection View region surveillance Alert event generation Event History Search & display

Foreground detection

View region surveillance

Alert event generation

Event History

Search & display

Contribution of SZTAKI Foreground detection in moving camera

Foreground detection in moving camera

Contribution of SZTAKI Mosaicing

Mosaicing

Contribution of SZTAKI Usual – non usual motion Pixel-wise motion estimation black: right, white: left Motion statistics Input Actual motion masked with usual motion

Usual – non usual motion

Contribution of SZTAKI SG based unusuality detector on motion fields motion tracks Software Environment Interface module to user dll /lib/module Separates and bridge modules Server Serves image/video streams Transcodes images Forward requests to modules DB server Metadata store & search Webserver Generate html pages with links to Server (later) Client dynamic web Javascript/flash based graphics display Mozilla native mjpeg stream + SVG

SG based unusuality detector on

motion fields

motion tracks

Software Environment

Interface module to user dll /lib/module

Separates and bridge modules

Server

Serves image/video streams

Transcodes images

Forward requests to modules

DB server

Metadata store & search

Webserver

Generate html pages with links to Server (later)

Client dynamic web

Javascript/flash based graphics display

Mozilla native mjpeg stream + SVG

Web Page DB metadata tcp/ip SERVER Web server Matlab C++ DLL/LIB Comm. Interface json tcp/ip tcp/ip mjpg json Html User modules Contribution of SZTAKI - architecture Data Source Controller Comm. Interface Comm. Interface Comm. Interface Module Register Streams Internet

Add a comment

Related presentations